Boundary-Compressed Approximate Pruning / bench_report.json

✓✓ Beats tuned baseline

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "grid": [
  9      {
 10        "config": {
 11          "lr": 0.0015,
 12          "epochs": 10,
 13          "keep": 16
 14        },
 15        "mean": 0.718140721321106,
 16        "per_seed": [
 17          1.2147727012634277,
 18          0.9435746669769287,
 19          0.3435657322406769,
 20          0.3706497848033905
 21        ]
 22      },
 23      {
 24        "config": {
 25          "lr": 0.003,
 26          "epochs": 10,
 27          "keep": 16
 28        },
 29        "mean": 0.759648285806179,
 30        "per_seed": [
 31          1.2128480672836304,
 32          1.0560415983200073,
 33          0.36093801259994507,
 34          0.4087654650211334
 35        ]
 36      },
 37      {
 38        "config": {
 39          "lr": 0.006,
 40          "epochs": 10,
 41          "keep": 16
 42        },
 43        "mean": 0.8399795442819595,
 44        "per_seed": [
 45          1.2241541147232056,
 46          0.9892252683639526,
 47          0.5402792096138,
 48          0.6062595844268799
 49        ]
 50      }
 51    ],
 52    "best": {
 53      "lr": 0.0015,
 54      "epochs": 10,
 55      "keep": 16
 56    },
 57    "full": {
 58      "per_seed": [
 59        1.2147727012634277,
 60        0.9435746669769287,
 61        0.3435657322406769,
 62        0.3706497848033905,
 63        0.942135751247406,
 64        0.7555773258209229,
 65        1.2186181545257568,
 66        0.6535401940345764
 67      ]
 68    }
 69  },
 70  "idea": {
 71    "config": {
 72      "lr": 0.0015,
 73      "epochs": 10,
 74      "eta": 2,
 75      "keep": 16
 76    },
 77    "per_seed": [
 78      0.9471153020858765,
 79      0.7885468602180481,
 80      0.3400212824344635,
 81      0.34854480624198914,
 82      0.5234649777412415,
 83      0.3543352484703064,
 84      0.9228858947753906,
 85      0.36787140369415283
 86    ],
 87    "details": [
 88      {
 89        "seed": 0,
 90        "metric": 0.9471153020858765,
 91        "keep": 16,
 92        "peak_states": 64,
 93        "graph_edges": 61,
 94        "retained_fraction": 0.5,
 95        "history_last": 0.8118170817693074
 96      },
 97      {
 98        "seed": 1,
 99        "metric": 0.7885468602180481,
100        "keep": 16,
101        "peak_states": 64,
102        "graph_edges": 61,
103        "retained_fraction": 0.5,
104        "history_last": 0.8393829234441121
105      },
106      {
107        "seed": 2,
108        "metric": 0.3400212824344635,
109        "keep": 16,
110        "peak_states": 64,
111        "graph_edges": 61,
112        "retained_fraction": 0.5,
113        "history_last": 0.370681676864624
114      },
115      {
116        "seed": 3,
117        "metric": 0.34854480624198914,
118        "keep": 16,
119        "peak_states": 64,
120        "graph_edges": 61,
121        "retained_fraction": 0.5,
122        "history_last": 0.3139439268906911
123      },
124      {
125        "seed": 4,
126        "metric": 0.5234649777412415,
127        "keep": 16,
128        "peak_states": 64,
129        "graph_edges": 61,
130        "retained_fraction": 0.5,
131        "history_last": 0.48530861457188923
132      },
133      {
134        "seed": 5,
135        "metric": 0.3543352484703064,
136        "keep": 16,
137        "peak_states": 64,
138        "graph_edges": 61,
139        "retained_fraction": 0.5,
140        "history_last": 0.335369356473287
141      },
142      {
143        "seed": 6,
144        "metric": 0.9228858947753906,
145        "keep": 16,
146        "peak_states": 64,
147        "graph_edges": 61,
148        "retained_fraction": 0.5,
149        "history_last": 0.8587548828125
150      },
151      {
152        "seed": 7,
153        "metric": 0.36787140369415283,
154        "keep": 16,
155        "peak_states": 64,
156        "graph_edges": 61,
157        "retained_fraction": 0.5,
158        "history_last": 0.3451614006360372
159      }
160    ]
161  },
162  "comparison": {
163    "delta_mean": -0.23120606690645218,
164    "idea_wins": 8,
165    "n_pairs": 8,
166    "per_seed_diffs": [
167      -0.26765739917755127,
168      -0.15502780675888062,
169      -0.003544449806213379,
170      -0.022104978561401367,
171      -0.41867077350616455,
172      -0.40124207735061646,
173      -0.2957322597503662,
174      -0.2856687903404236
175    ],
176    "p_value": 0.0081,
177    "mde": 0.13177997997997729,
178    "mde_rel_pct": 16.363998280184262,
179    "verdict": "idea better (significant)",
180    "system_worked": true
181  },
182  "mechanism_signature": {
183    "prediction": "omitted-token interaction mass predicts output sensitivity",
184    "predicted_mean": 0.09890676662325859,
185    "observed_mean": 0.06862932164222002,
186    "correlation": -0.32015715855609533,
187    "confirmed": false
188  },
189  "idea_sweep": [
190    {
191      "config": {
192        "lr": 0.0015,
193        "epochs": 10,
194        "eta": 2,
195        "keep": 16
196      },
197      "mean": 0.5740982219576836,
198      "per_seed": [
199        0.9471153020858765,
200        0.7885468602180481,
201        0.3400212824344635,
202        0.34854480624198914,
203        0.5234649777412415,
204        0.3543352484703064,
205        0.9228858947753906,
206        0.36787140369415283
207      ]
208    },
209    {
210      "config": {
211        "lr": 0.003,
212        "epochs": 10,
213        "eta": 2,
214        "keep": 16
215      },
216      "mean": 0.5850291326642036,
217      "per_seed": [
218        0.970460057258606,
219        0.8036816716194153,
220        0.3443853259086609,
221        0.37667813897132874,
222        0.5293618440628052,
223        0.35860389471054077,
224        0.9097486138343811,
225        0.38731351494789124
226      ]
227    },
228    {
229      "config": {
230        "lr": 0.006,
231        "epochs": 10,
232        "eta": 2,
233        "keep": 16
234      },
235      "mean": 0.6615562364459038,
236      "per_seed": [
237        1.0524933338165283,
238        1.077573299407959,
239        0.3749794065952301,
240        0.4042664170265198,
241        0.5629213452339172,
242        0.3612123727798462,
243        1.0209476947784424,
244        0.43805602192878723
245      ]
246    }
247  ],
248  "runtime_sec": 65.80087637901306,
249  "structural_match": "multi-token sequence forecasting with token interaction graph",
250  "state_counts": [
251    64,
252    64,
253    64,
254    64,
255    64,
256    64,
257    64,
258    64
259  ]
260}